WIP: Text to SQL query

chatgpt-plugin/src/pages/api/handlers/generic-sql/examples.ts

@@ -0,0 +1,166 @@
+const examplesText2Sql = [
+  // Domain-specific
+  {
+    user: "Most expensive NFT sale",
+    assistant:
+      "SELECT seller, mint, sales_amount FROM solana.core.fact_nft_sales WHERE sales_amount IS NOT NULL ORDER BY sales_amount DESC LIMIT 1;",
+  },
+  {
+    user: "Mint price and purchaser of the 10 latest NFT mints",
+    assistant:
+      "SELECT mint, mint_price, PURCHASER FROM solana.core.fact_nft_mints DESC LIMIT 10;",
+  },
+  {
+    user: "What are the highest priced NFTs in the last 30 days?",
+    assistant:
+      "SELECT mint, MAX(sales_amount) as max_nft_price FROM solana.core.fact_nft_sales WHERE block_timestamp > CURRENT_DATE - interval '30 days';",
+  },
+  //
+  // WITH yesterday_sales AS (
+  //     SELECT *
+  //     FROM solana.core.fact_nft_sales
+  //     WHERE
+  //       block_timestamp >= CURRENT_DATE - interval '1 day'
+  //       AND block_timestamp < CURRENT_DATE
+  //       AND sales_amount IS NOT NULL
+  // )
+  // SELECT seller, mint, sales_amount
+  // FROM yesterday_sales
+  // ORDER BY sales_amount DESC;
+  {
+    user: "NFT sales yesterday",
+    assistant:
+      "WITH yesterday_sales AS( SELECT * FROM solana.core.fact_nft_sales WHERE block_timestamp >= CURRENT_DATE - interval '1 day' AND block_timestamp < CURRENT_DATE AND sales_amount IS NOT NULL) SELECT seller, mint, sales_amount FROM yesterday_sales ORDER BY sales_amount DESC;",
+  },
+  //
+  // SELECT seller, sales_amount, block_timestamp
+  // FROM (
+  // SELECT seller, sales_amount, fact_nft_sales.block_timestamp
+  // FROM solana.core.fact_nft_sales as fact_nft_sales
+  // INNER JOIN solana.core.fact_nft_mints as fact_nft_mints
+  // ON fact_nft_sales.mint = fact_nft_mints.mint
+  // WHERE fact_nft_sales.mint = ANY(
+  //     SELECT mint
+  //     FROM solana.core.dim_nft_metadata
+  //     WHERE project_name = 'Solana Monkey Business'
+  // )
+  // AND fact_nft_sales.sales_amount IS NOT NULL
+  // )
+  // ORDER BY sales_amount DESC
+  // LIMIT 10;
+  {
+    user: "Top sales for Solana Monkey Business",
+    assistant:
+      "SELECT seller, sales_amount, block_timestamp FROM( SELECT seller, sales_amount, fact_nft_sales.block_timestamp FROM solana.core.fact_nft_sales as fact_nft_sales INNER JOIN solana.core.fact_nft_mints as fact_nft_mints ON fact_nft_sales.mint = fact_nft_mints.mint WHERE fact_nft_sales.mint = ANY( SELECT mint FROM solana.core.dim_nft_metadata WHERE project_name = 'Solana Monkey Business') AND fact_nft_sales.sales_amount IS NOT NULL ) ORDER BY sales_amount DESC LIMIT 10;",
+  },
+  //
+  // SELECT seller, mint, sales_amount
+  // FROM solana.core.fact_nft_sales
+  // WHERE
+  // sales_amount = (
+  //     SELECT MAX(sales_amount)
+  //     FROM solana.core.fact_nft_sales
+  //     WHERE block_timestamp > CURRENT_DATE - interval '30 days' AND sales_amount IS NOT NULL
+  // )
+  // ORDER BY block_timestamp DESC
+  // LIMIT 1;
+  {
+    user: "Who sold the most expensive NFT in the last 30 days? What was the price?",
+    assistant:
+      "SELECT seller, mint, sales_amount FROM solana.core.fact_nft_sales WHERE sales_amount =( SELECT MAX(sales_amount) FROM solana.core.fact_nft_sales WHERE block_timestamp > CURRENT_DATE - interval '30 days' AND sales_amount IS NOT NULL) ORDER BY block_timestamp DESC LIMIT 1;",
+  },
+  //
+  // SELECT purchaser, SUM(sales_amount) as spent_amount
+  // FROM solana.core.fact_nft_sales
+  // WHERE block_timestamp > CURRENT_DATE - interval '7 days' AND sales_amount IS NOT NULL
+  // GROUP BY purchaser
+  // HAVING spent_amount > 100
+  // ORDER BY spent_amount DESC;
+  {
+    user: "Who spend more than 100 SOL on buying NFTs last week?",
+    assistant:
+      "SELECT purchaser, SUM(sales_amount) as spent_amount FROM solana.core.fact_nft_sales WHERE block_timestamp > CURRENT_DATE - interval '7 days' AND sales_amount IS NOT NULL GROUP BY purchaser HAVING spent_amount > 100 ORDER BY spent_amount DESC;",
+  },
+  // SELECT seller, fact_nft_sales.mint, (fact_nft_sales.sales_amount - fact_nft_mints.mint_price) as price_difference
+  // FROM solana.core.fact_nft_sales as fact_nft_sales
+  // INNER JOIN solana.core.fact_nft_mints fact_nft_mints
+  // ON fact_nft_sales.seller = fact_nft_mints.purchaser AND fact_nft_sales.sales_amount IS NOT NULL
+  // ORDER BY price_difference DESC
+  // LIMIT 3;
+  {
+    user: "What were the three highest NFT sales fron mint price?",
+    assistant:
+      "SELECT seller, fact_nft_sales.mint,(fact_nft_sales.sales_amount - fact_nft_mints.mint_price) as price_difference FROM solana.core.fact_nft_sales as fact_nft_sales INNER JOIN solana.core.fact_nft_mints fact_nft_mints ON fact_nft_sales.seller = fact_nft_mints.purchaser AND fact_nft_sales.sales_amount IS NOT NULL ORDER BY price_difference DESC LIMIT 3;",
+  },
+  //
+  // SELECT
+  //     seller,
+  //     fact_nft_sales.mint,
+  //     (fact_nft_sales.sales_amount - fact_nft_mints.mint_price) as flip_profit
+  // FROM solana.core.fact_nft_sales as fact_nft_sales
+  // INNER JOIN solana.core.fact_nft_mints fact_nft_mints
+  // ON
+  //     fact_nft_sales.seller = fact_nft_mints.purchaser
+  //     AND fact_nft_sales.sales_amount IS NOT NULL
+  //     AND fact_nft_sales.block_timestamp < fact_nft_mints.block_timestamp + interval '1 day'
+  // ORDER BY flip_profit DESC
+  // LIMIT 1;
+  {
+    user: "What is the most profitable flip within the day of minting an NFT?",
+    assistant:
+      "SELECT seller, fact_nft_sales.mint,(fact_nft_sales.sales_amount - fact_nft_mints.mint_price) as flip_profit FROM solana.core.fact_nft_sales as fact_nft_sales INNER JOIN solana.core.fact_nft_mints fact_nft_mints ON fact_nft_sales.seller = fact_nft_mints.purchaser AND fact_nft_sales.sales_amount IS NOT NULL AND fact_nft_sales.block_timestamp < fact_nft_mints.block_timestamp + interval '1 day' ORDER BY flip_profit DESC LIMIT 1;",
+  },
+  // Generic
+  // {
+  //   user: "What zip code has the highest percentage of people of age 75?",
+  //   assistant:
+  //     "SELECT location_data.zip_code, \\n       (population_by_age.population_75_to_84_years / total_population.total_population) * 100 AS percentage_75_to_84\\nFROM population_by_age\\nJOIN location_data ON population_by_age.geo_id = location_data.geo_id\\nJOIN total_population ON population_by_age.geo_id = total_population.geo_id\\nWHERE total_population.total_population > 0\\nORDER BY percentage_75_to_84 DESC NULLS LAST\\nLIMIT 1;",
+  // },
+  // {
+  //   user: "Which 5 counties have the most arson?",
+  //   assistant:
+  //     "SELECT location_data.county, location_data.state, SUM(crime_by_city.arson) AS total_arson\\nFROM crime_by_city\\nJOIN location_data ON crime_by_city.city = location_data.city AND crime_by_city.state = location_data.state\\nWHERE crime_by_city.arson IS NOT NULL\\nGROUP BY location_data.county\\nORDER BY total_arson DESC NULLS LAST\\nLIMIT 5;",
+  // },
+  // {
+  //   user: "Which 5 cities have the most females?",
+  //   assistant:
+  //     "SELECT location_data.city, location_data.state, population_by_gender.female_population\\nFROM location_data\\nJOIN population_by_gender ON location_data.geo_id = population_by_gender.geo_id\\nORDER BY population_by_gender.female_population DESC NULLS LAST\\nLIMIT 5;",
+  // },
+  // {
+  //   user: "Which city in Washington has the highest population?",
+  //   assistant:
+  //     "SELECT location_data.city, location_data.state, \\n       SUM(population_by_gender.male_population) + SUM(population_by_gender.female_population) AS total_population\\nFROM location_data\\nJOIN population_by_gender ON location_data.geo_id = population_by_gender.geo_id\\nWHERE location_data.state = 'WA'\\nGROUP BY location_data.city, location_data.state\\nORDER BY total_population DESC NULLS LAST\\nLIMIT 1;",
+  // },
+  // {
+  //   user: "Which zip code in San Francisco has the highest racial diversity and what is the percentage population of each race in that zip code?",
+  //   assistant:
+  //     "SELECT location_data.zip_code, \\n       (population_by_race.white_population / NULLIF(total_population.total_population, 0)) * 100 AS percentage_white,\\n       (population_by_race.black_population / NULLIF(total_population.total_population, 0)) * 100 AS percentage_black,\\n       (population_by_race.native_american_population / NULLIF(total_population.total_population, 0)) * 100 AS percentage_native_american,\\n       (population_by_race.asian_population / NULLIF(total_population.total_population, 0)) * 100 AS percentage_asian,\\n       (population_by_race.hispanic_population / NULLIF(total_population.total_population, 0)) * 100 AS percentage_hispanic,\\n       (population_by_race.two_or_more_population / NULLIF(total_population.total_population, 0)) * 100 AS percentage_two_or_more\\nFROM population_by_race\\nJOIN location_data ON population_by_race.geo_id = location_data.geo_id\\nJOIN (SELECT geo_id, SUM(white_population + black_population + native_american_population + asian_population + hispanic_population + two_or_more_population) AS total_population\\n      FROM population_by_race\\n      GROUP BY geo_id) AS total_population ON population_by_race.geo_id = total_population.geo_id\\nWHERE location_data.city = 'San Francisco' AND location_data.state = 'CA'\\nORDER BY (population_by_race.white_population + population_by_race.black_population + population_by_race.native_american_population + population_by_race.asian_population + population_by_race.hispanic_population + population_by_race.two_or_more_population) DESC NULLS LAST\\nLIMIT 1;",
+  // },
+  // {
+  //   user: "Zip code in California with the most advanced degree holders",
+  //   assistant:
+  //     "SELECT location_data.zip_code, population_by_education_level.masters_degree + population_by_education_level.professional_school_degree + population_by_education_level.doctorate_degree AS total_advanced_degrees\\nFROM population_by_education_level\\nJOIN location_data ON population_by_education_level.geo_id = location_data.geo_id\\nWHERE location_data.state = 'CA'\\nORDER BY total_advanced_degrees DESC NULLS LAST\\nLIMIT 1;",
+  // },
+];
+
+const examplesTableSelection = [
+  {
+    user: "Which top 5 cities have the most total crime?",
+    assistant: '\n{\n    "tables": ["crime_by_city"]\n}\n',
+  },
+  {
+    user: "What zip code has the highest percentage of people of age 75 and over?",
+    assistant: '\n{\n    "tables": ["location_data", "population_by_age"]\n}\n',
+  },
+  {
+    user: "Which 5 counties have the most arson?",
+    assistant: '\n{\n    "tables": ["crime_by_city", "location_data"]\n}\n',
+  },
+  {
+    user: "Which city has the most total crime and what is the racial makeup of that city?",
+    assistant:
+      '\n{\n    "tables": ["crime_by_city", "location_data", "population_by_race"]\n}\n',
+  },
+];
+
+export { examplesText2Sql, examplesTableSelection };

chatgpt-plugin/src/pages/api/handlers/generic-sql/index.ts

@@ -0,0 +1,168 @@
+import { Flipside, Query, QueryResultSet } from "@flipsidecrypto/sdk";
+import {
+  ChatCompletionResponseMessage,
+  Configuration,
+  OpenAIApi,
+} from "openai";
+import { NextApiRequest, NextApiResponse } from "next";
+import { encoding_for_model } from "tiktoken";
+import type { TiktokenModel } from "tiktoken";
+import * as util from "util";
+
+import {
+  getText2SqlExampleMessages,
+  enrichPromptForFewShotInference,
+} from "./prompt";
+import type { IMessage } from "./prompt";
+import tablesCore from "./tables-flipside";
+
+const modelName: TiktokenModel = "text-davinci-003";
+const DIALECT = "PostgreSQL 15.2"; // or 'MySQL'
+
+function getTablesInMachinePromptForm(tableNames: string[]): string {
+  return tablesCore
+    .map((table) => {
+      const tableName = `table name: ${table.name}`;
+      const tableDescription = `table description: ${table.description}`;
+      const tableColumns = table.columns.map(
+        (column) => `${column.name} ${column.type} ${column.description}`
+      );
+      return [tableName, tableDescription, ...tableColumns].join("\n");
+    })
+    .join("\n\n");
+}
+
+async function getAssistantMessageFromOpenAi(
+  messages: IMessage[],
+  modelName: TiktokenModel
+): Promise<ChatCompletionResponseMessage> {
+  const openai = new OpenAIApi(
+    new Configuration({
+      apiKey: process.env.OPENAI_API_KEY,
+    })
+  );
+
+  const completion = await openai.createChatCompletion({
+    // model: modelName,
+    // https://platform.openai.com/docs/models/model-endpoint-compatibility
+    // Available: gpt-4, gpt-4-0314, gpt-4-32k, gpt-4-32k-0314, gpt-3.5-turbo, gpt-3.5-turbo-0301
+    model: "gpt-3.5-turbo",
+    messages,
+    temperature: 0,
+  });
+
+  if (completion.data.choices.length === 0)
+    throw new Error("No response from OpenAI");
+  return completion.data.choices[0].message as ChatCompletionResponseMessage;
+}
+
+// Takes a natural language prompt and try to generate a valid SQL to answer it `retries` times
+async function promptToSqlWithRetry(
+  prompt: string,
+  retries = 3
+): Promise<string> {
+  const encoder = encoding_for_model(modelName);
+
+  const tableNames = tablesCore.map((table) => table.name);
+  const tablePrompt = getTablesInMachinePromptForm(tableNames);
+
+  const schemaMessage: IMessage[] = [
+    { role: "user", content: tablePrompt },
+  ];
+
+  const instLen = encoder.encode(tablePrompt).length; // Instruction Length
+  const exampleMessages = getText2SqlExampleMessages(instLen);
+  console.log('exampleMessages', exampleMessages)
+
+  const messageHistory: IMessage[] = [
+    { role: "user", content: enrichPromptForFewShotInference(prompt, DIALECT) },
+  ];
+  console.log('messageHistory', messageHistory)
+
+  let assistantMessage: ChatCompletionResponseMessage;
+  let sqlQuery = "";
+
+  // Try to generate SQL query from prompt for `retries` times
+  for (let k = 0; k < retries; k++) {
+    try {
+      try {
+        let payload = schemaMessage.concat(messageHistory);
+        if (k === 0) payload = exampleMessages.concat(payload);
+        assistantMessage = await getAssistantMessageFromOpenAi(
+          payload,
+          modelName
+        );
+
+        // Try to convert OpenAI's output to JSON and extract `SQL` field.
+        // If it fails, try again. Otherwise, use the SQL query.
+        sqlQuery = JSON.parse(assistantMessage.content).SQL as string;
+        if (sqlQuery) break;
+      } catch (e) {
+        // ignore, don't handle error
+        console.log((e as any).toJSON());
+      }
+    } catch (e) {
+      console.log(e);
+      throw e;
+    }
+  }
+
+  if (sqlQuery === "") {
+    throw new Error("Could not generate SQL query from prompt");
+  }
+
+  // release encoder
+  encoder.free();
+  return sqlQuery;
+}
+
+async function dispatchSqlQueryFlipside(
+  sqlQuery: string
+): Promise<QueryResultSet | undefined> {
+  try {
+    // Initialize `Flipside` with your API key
+    const flipside = new Flipside(
+      process.env.FLIPSIDE_CRYPTO_API_KEY as string,
+      "https://api-v2.flipsidecrypto.xyz"
+    );
+
+    const query: Query = {
+      sql: sqlQuery,
+      maxAgeMinutes: 1,
+    };
+
+    return flipside.query.run(query);
+  } catch (error) {
+    console.log(util.inspect((error as any).response.data, false, null, true));
+  }
+}
+
+export default async function handler(
+  req: NextApiRequest,
+  res: NextApiResponse
+) {
+  if (req.method != "POST") {
+    res.status(405).send({ message: "Only POST requests allowed" });
+    return;
+  }
+
+  let prompt = req.body.prompt;
+  if (!prompt) {
+    res.status(400).send({ message: "Missing prompt" });
+    return;
+  }
+
+  try {
+    const sqlQuery = await promptToSqlWithRetry(prompt, 3);
+    // console.log("sqlQuery", sqlQuery);
+
+    const data = await dispatchSqlQueryFlipside(sqlQuery);
+    // console.log(data);
+
+    res.status(200).send({ message: data });
+  } catch (e) {
+    console.log(e);
+    res.status(500).send({ error: true, message: "Internal server error" });
+    return;
+  }
+}

chatgpt-plugin/src/pages/api/handlers/generic-sql/prompt.ts

@@ -0,0 +1,52 @@
+import { v4 as uuidv4 } from "uuid";
+
+import { examplesText2Sql } from "./examples";
+
+// FROM: https://github.com/caesarHQ/textSQL/blob/main/api/app/api/utils/sql_gen/prompts.py#L145
+// Crates the retry prompt for the given scope formatted to the given dialect and schemas.
+export function enrichPromptForFewShotInference(prompt: string, dialect: string): string {
+  return `
+generation_id: ${uuidv4()}
+You are an expert and empathetic database engineer that is generating correct read-only ${dialect} query to answer the following question/command: ${prompt}
+
+- Use state abbreviations for states.
+- Table crime_by_city does not have columns 'zip_code' or 'county'.
+- Do not use ambiguous column names.
+- For example, city can be ambiguous because both tables location_data and crime_by_city have a column named city. Always specify the table where you are using the column.
+- If you include a city or county column in the result table, include a state column too.
+- Make sure each value in the result table is not null.
+- Use CTE format for computing subqueries.
+
+Provide a properly formatted JSON object with the following information. Ensure to escape any special characters (e.g. \n should be \\n, \m \\m and such) so it can be parsed as JSON.
+{{
+    "Schema": "<1 to 2 sentences about the tables/columns/enums above to use>",
+    "SQL": "<your query>"
+}}
+
+Command: ${prompt}
+`;
+}
+
+export interface IMessage {
+  role: "user" | "assistant";
+  content: string;
+}
+
+export function getText2SqlExampleMessages(instLen: number): IMessage[] {
+  // const maxMessages = instLen > 1000 ? 5 : instLen > 1500 ? 2 : instLen > 2000 ? 1 : -1;
+
+  // For now, just use all examples
+  let examples = examplesText2Sql;
+
+  // if (maxMessages > 0) {
+  //   examples = examplesText2Sql.slice(0, maxMessages);
+  // }
+
+  const messages: IMessage[] = [];
+  for (const example of examples) {
+    messages.push({ role: "user", content: example.user });
+    messages.push({ role: "assistant", content: example.assistant });
+  }
+
+  return messages;
+}